Amazon SageMaker HyperPod Now Lets You Choose the Best Network Topology per Partition!

Hey everyone, I'm Shii-chan! Today I want to share an update to Amazon SageMaker HyperPod — great news if you run large GPU clusters!

AWS What's New aws.amazon.com

What was announced?

According to AWS's What's New, Amazon SageMaker HyperPod now supports network topology configuration at the partition level for Slurm-orchestrated clusters!

A single cluster can now mix topologies: one partition can run tree topology while another runs block topology. Each partition can use whichever topology best matches its instance types, which helps boost distributed training performance.

The story so far

Until now, a cluster could only use a single topology for the whole thing. But real-world clusters often mix different instance types, and running a topology that doesn't match an instance type's interconnect characteristics could hurt GPU-to-GPU communication efficiency.

What changes

HyperPod now looks at each partition's instance types and automatically determines and maintains the right topology. Job placement lines up with each instance type's interconnect characteristics, so GPU-to-GPU communication gets faster, NCCL collective operations run more efficiently, and training throughput improves.

Best of all, there's nothing to configure — topology-aware scheduling is enabled by default! If your team runs a large GPU cluster, you get the benefit automatically.

Dive Deep

Here's how topology assignment works:

• Block topology: partitions using Amazon EC2 UltraServer instance types (e.g., ml.p6e-gb200.36xlarge)
• Tree topology: partitions using hierarchical-interconnect instance types (e.g., ml.p5.48xlarge, ml.p5e.48xlarge, ml.p5en.48xlarge)
• Partitions with instance types that don't expose network topology information stay fully schedulable

HyperPod keeps this configuration up to date automatically as the cluster changes — through scale-up, scale-down, and node replacement — so each partition's topology always reflects the cluster's current state.

To get started, create a new SageMaker HyperPod Slurm cluster (or update an existing one) running Slurm 25.11 or later with supported GPU instance types. The feature is available in every AWS Region where Amazon SageMaker HyperPod is supported.

Wrap-up

• SageMaker HyperPod Slurm clusters can now mix tree and block network topologies across partitions
• HyperPod automatically determines and maintains the right topology per partition based on instance types, even through scaling and node replacement
• No configuration needed — it's on by default. You just need Slurm 25.11+ and supported GPU instance types
• A solid win for engineers running large-scale distributed training on HyperPod, boosting communication efficiency and throughput